---
license: apache-2.0
language: en
base_model: google/flan-t5-base
datasets: [sudeep1610/resumebot-dataset]
library_name: transformers
tags: [flan-t5, fine-tuned, resume, chatbot, rag, retrieval-augmented-generation, gradio, personal-assistant]
---

# 🤖 ResumeBot: a personalised LLM fine-tuned on my resume

**ResumeBot** is a chatbot that answers questions about **Sudeep** (Data Analyst | Data Science).
It is Google's pretrained **google/flan-t5-base** model **fine-tuned on 353 question–answer pairs built from my resume**,
combined with a semantic retriever and two hallucination guards so it only says things that are actually on my resume.

> *"What are Sudeep's technical skills?"* → *"Sudeep's technical skills are Python, SQL, Power BI, Excel, Data Analysis,
> Data Visualization and Business Intelligence."* 📄 *Resume section: Technical skills*

## How it works

```
Question
   ↓
Semantic retriever (all-MiniLM-L6-v2) finds the most relevant resume facts
   ↓
Out-of-scope guard: not about the resume?  →  polite refusal (no guessing)
   ↓
Fine-tuned FLAN-T5 writes the answer from those facts
   ↓
Grounding check: answer contains words not in the resume?  →  return the verified resume fact
   ↓
Answer + resume section (Gradio web app)
```

This approach is called **retrieval-augmented fine-tuning**: the model is fine-tuned to answer from retrieved resume text,
which keeps answers accurate even for questions worded in new ways.

## Training details

| | |
|---|---|
| Base model | [`google/flan-t5-base`](https://huggingface.co/google/flan-t5-base) (pretrained by Google) |
| Method | Supervised fine-tuning (seq2seq) with retrieved context |
| Knowledge base | 28 resume facts |
| Dataset | [sudeep1610/resumebot-dataset](https://huggingface.co/datasets/sudeep1610/resumebot-dataset): 353 Q&A pairs (327 train / 26 unseen test) |
| Epochs / learning rate / batch size | 10 / 0.0003 / 8 |
| Hardware | Google Colab (free T4 GPU) |
| Retriever | sentence-transformers/all-MiniLM-L6-v2 |

## Results (on questions never seen during training)

| Metric | Value |
|---|---|
| Final training loss | 0.0010 |
| Loss on unseen questions | 0.0056 |
| Retriever picked the correct resume fact | 69% |
| Answer exactly matches the resume answer | 69% |

<img src="loss_curve.png" alt="Training loss" width="620">

## Hallucination guards
1. **Retrieval:** the model only sees the resume facts relevant to the question.
2. **Grounding check:** any answer containing information that is not in the retrieved resume text is replaced by the verified fact.
3. **Out-of-scope guard:** questions unrelated to the resume (e.g. *"What is the capital of France?"*) get a polite refusal.

## Use it

**Run the full ResumeBot web app (Gradio):**
```bash
git clone https://huggingface.co/sudeep1610/ResumeBot
cd ResumeBot
pip install -r requirements.txt
python app.py
```

**Use only the fine-tuned model in Python:**
```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tok = AutoTokenizer.from_pretrained("sudeep1610/ResumeBot")
model = AutoModelForSeq2SeqLM.from_pretrained("sudeep1610/ResumeBot")
prompt = ("Resume information: Sudeep's technical skills are Python, SQL, Power BI, Excel, Data Analysis, "
          "Data Visualization and Business Intelligence.\nQuestion: What tools does he know?\n"
          "Answer using only the resume information:")
out = model.generate(**tok(prompt, return_tensors="pt"), max_new_tokens=128, num_beams=4)
print(tok.decode(out[0], skip_special_tokens=True))
```

## Files

| File | What it is |
|---|---|
| `model.safetensors`, `config.json`, `tokenizer*` | The fine-tuned FLAN-T5 model |
| `app.py` | Complete ResumeBot: retriever + fine-tuned model + guards + Gradio website |
| `bot_config.json` | Knowledge base: resume facts, profile, prompt, known questions |
| `requirements.txt` | Python packages needed to run `app.py` |
| `ResumeBot_Pro.ipynb` | Training notebook (Google Colab) |
| `loss_curve.png` | Training / test loss chart |

## Limitations
- Knows only what is written in the resume; it cannot answer about anything else by design.
- Small academic project: trained on a small dataset, so unusual phrasings may occasionally match the wrong section.

## Author
**Sudeep**, Data Analyst | Data Science · M.Sc. Data Science, Dayananda Sagar University (2025-2027)
[LinkedIn](https://linkedin.com/in/sudeep-shetty-0aa23a1b0) · [GitHub](https://github.com/sudeepshetty4247-stack)